The biggest week in AI (GPT-4, Office Copilot, Google PaLM, Anthropic Claude & more)

#mlnews #gpt4 #copilot
Your weekly news all around the AI world
Check out W&B courses (free): wandb.courses/
OUTLINE:
0:00 - Intro
0:20 - GPT-4 announced!
4:30 - GigaGAN: The comeback of Generative Adversarial Networks
7:55 - ChoppedAI: AI Recipes
8:45 - Samsung accused of faking space zoom effect
14:00 - Weights & Biases courses are free
16:55 - Data Portraits
18:50 - Data2Vec 2.0
19:50 - Gated Models on Hugging Face & huggingface.js
22:05 - Visual ChatGPT
23:35 - Bing crosses 100 million daily active users
24:50 - Casual Conversations Dataset
25:50 - Anthropic AI Safety Research
27:30 - Magnushammer & more advances in AI-assisted math
30:30 - LLaMA license change PR
32:00 - Self-Instruct dataset
33:35 - PaLM-E: Multimodal Pathways
35:45 - USM: Universal Speech Model
37:25 - GILGEN: Grounded Text-to-Image
39:55 - Fruit Fly Connectome released
References:
www.heise.de/news/GPT-4-kommt...
mingukkang.github.io/GigaGAN/
www.choppedai.com/
/ samsung_space_zoom_moo...
imgur.com/ULVX933
imgur.com/9XMgt06
imgur.com/9kichAp
imgur.com/RSHAz1l
imgur.com/PIAjVKp
imgur.com/xEyLajW
imgur.com/3STX9mZ
imgur.com/ifIHr3S
imgur.com/bXJOZgI
dataportraits.org/
arxiv.org/abs/2303.03919
arxiv.org/pdf/2303.03919.pdf
/ ai-self-supervised-lea...
github.com/facebookresearch/f...
huggingface.co/docs/hub/model...
huggingface.co/about
github.com/huggingface/huggin...
github.com/microsoft/visual-c...
arxiv.org/abs/2303.04671
github.com/microsoft/visual-c...
huggingface.co/spaces/RamAnan...
www.engadget.com/microsoft-bi...
/ casual-conversations-v...
/ casual-conversations-v...
www.anthropic.com/index/core-...
arxiv.org/abs/2303.04488
arxiv.org/pdf/2303.04488.pdf
arxiv.org/abs/2303.04910
arxiv.org/pdf/2303.04910.pdf
/ 1633645134934949888
ai.papers.bar/paper/ede58b1eb...
arxiv.org/pdf/2303.03192.pdf
www.theverge.com/2023/3/8/236...
knightcolumbia.org/blog/the-l...
github.com/facebookresearch/l...
huggingface.co/datasets/yizho...
openai.com/policies/terms-of-use
palm-e.github.io/
pickapic.io/
ai.googleblog.com/2023/03/uni...
arxiv.org/abs/2303.01037
github.com/BlinkDL/RWKV-LM?ut...
gligen.github.io/
github.com/microsoft/GLIP
arxiv.org/abs/2301.07093
huggingface.co/spaces/gligen/...
www.sciencealert.com/the-firs...
en.wikipedia.org/wiki/Tidal_l...
Links:
Homepage: ykilcher.com
Merch: ykilcher.com/merch
KZbin: / yannickilcher
Twitter: / ykilcher
Discord: ykilcher.com/discord
LinkedIn: / ykilcher
If you want to support me, the best thing to do is to share out the content :)
If you want to support me financially (completely optional and voluntary, but a lot of people have asked for this):
SubscribeStar: www.subscribestar.com/yannick...
Patreon: / yannickilcher
Bitcoin (BTC): bc1q49lsw3q325tr58ygf8sudx2dqfguclvngvy2cq
Ethereum (ETH): 0x7ad3513E3B8f66799f507Aa7874b1B0eBC7F85e2
Litecoin (LTC): LQW2TRyKYetVC8WjFkhpPhtpbDM4Vw7r9m
Monero (XMR): 4ACL8AGrEo5hAir8A9CeVrW8pEauWvnp1WnSDZxW7tziCDLhZAGsgzhRQABDnFy8yuM9fWJDviJPHKRjV4FWt19CJZN9D4n

[@gJonii]

I love this blast from the past, reminding us of the simpler times, times like last Monday.

[@y__h]

"What a year huh?" "Captain it's March"

[@gallacaecilia8009]

Hi Yannic, Silke Hahn here. I am the author of the original article you quote from Heise, who announced GPT-4's imminent launch. 🎉 As a tech journalist I focus on AI and published the news also in English: "GPT-4 is imminent, says Microsoft Germany - and it will be multimodal" ;) Anyways, we were all surprised that the CTO of Microsoft Germany came forward with this piece of information in front of a German business audience, before OpenAI themselves did. Four senior executives were present, amongst them the German CEO, and the whole session was dedicated to multimodality. Highly improbable that this was a slip of words (none of them spoke spontaneously, all talks had been pre-scheduled to the minute a week ahead and all questions were vetted, none taken from the audience on the spot). Therefore this was not an accidental announcement. We are left to wonder why in this odd context, but they explicitly spoke about "the iPhone moment" 😂 I had expected this launch any time between February/ March due to bits of information. That's why I recorded the session from scratch and was quite sure about the accuracy of my account. To me it was clear that GPT-4 had already been finished last year in autumn when OpenAI started negotiations with Microsoft about the next funding. And hey, frankly: those Hyperscaling platform salesmen know what they do, they wouldn't pay 10bn 💰 for a system that doesn't exist. Another thing: OpenAI sold its soul to Microsoft in 2019, before the launch of GPT-3, when they accepted the first billion in exchange for Microsoft's exclusive access. Ever since, they went closed source. I really hope for Open Source and European/ African/ etc international competitors. Monitoring the AI development with a critical eye because we are witnessing an emergent monopoly over a basic technology. We definitely need alternatives to that. That's why I was at that business event, apparently the only journalist on the spot. I was pretty sure that anybody of rank and file at Microsoft knows the state of the art of GPT-4 and hoped to hunt down some information. To me, there was no doubt their CTO provided us with reliable and accurate information. Microsoft did not deny the account and someone from their team got in touch via email, indirectly confirming the report. It wasn't meant to leak into the world at that time, I think, but that's what happened. Here we go, GPT-4 is out, Claude too and the year will surprise us with many more AI innovations than we have ever seen or imagined before. Also from small scale players. Exciting times! Thanks for the video, both entertaining and insightful. Next time you come across some implausible sounding news from a weird small country with a native language other than English, just get in touch with the author. As did The Verge and other media. Watch out and take care. ❤ Best from Germany Silke @_SilkeHahn 🐦

[@nocturnomedieval]

@Yannic_Kilcher

[@TiagoTiagoT]

We're approaching the point where news will only be able to be current by attempting to predict the future....

[@axolet]

OpenAi's TOS regarding traning data competition will FAIL. Why? Because the argument for AI training is that people are allowed to train their programs on art/writing/etc despite being competition to those very things. Therefore, if OpenAI successfully sues companies for training on their data, artists and writers can make the same argument that OpenAI is training with THEIR data to compete against them!

[@AltMarc]

In Germany there is already, some journalists and their unions whining/complaining about that their copyrighted and published texts being stolen by OpenAI and Co. They ask about changes in laws to get money of these AI firms...

[@HaldirZero]

The difference is that OpenAI and other corporations have tons of money for lawyers, and artists and authors don't.

[@axolotron1298]

I just saw(1) that Alpaca was designed following the Self-instruct technique. Lets see how that's gonna play out. If OpenAI waves its rights for the training set (maybe just for Alpaca) we could have one very good option for smaller uses. If they enforce the Terms of use, that could be also a big thing that could potentially damage MS business so they have to be careful with their decision. Still, Alpaca shows the effectiveness of the technique, so people can start building an open source dataset to fine tune Llama or others, by using Open Assistant, for instance. 1) kzbin.info/www/bejne/rqTPiGiphrSeeZo

[@drdca8263]

Were the things that OpenAI trained their models on, only available to people who agreed to a ToS which forbade using the data for training? It seems to me like it is possible that “don’t train on this” is something a ToS can require, but without copyright forbidding training on things.

[@unknownunknown5363]

@@AltMarc cuz they are poets. They believe in irrationalities much.

[@calvingrondahl1011]

Your GPT-4 intro correcting yourself was hilarious. 😊

[@BrutalStrike2]

I was expecting third clone comes and says it's released! 🎉

[@THEPAGMAN]

Thank you so much with doing these videos. I dont particularly have much time to go through the latest research as a student doing an AI masters, and these really help me stay up to date with the latest news!

[@majectic-berry]

Strong "Bitter Lesson" vibes from that GigaGan announcement... As is tradition nowadays lol

[@jddes]

Thanks for keeping us updated on the news! I'm very interested to hear your take on Alpaca, though.

[@blubberkumpel6740]

you are the best news channel for me atm. Thank you so much!

[@AdamRaudonis]

Another fantastic video! Thank you!

[@vivienseguy]

These ML news are so thorough, it's amazing.

[@joseph7858]

thank you, Yannic, for your entertaining summary! 😊

[@xxdaggerxx5]

45mins aint a summary

[@electric_sand]

@@xxdaggerxx5 It's a combination of summaries

[@SupremeCobraCommander]

Love that intro! I knew I recognized you from somewhere!

[@ImaplanetJupiteeeerr]

Nice to see a new ML news video after awhile! ^^

[@shinkurt]

Thanks yan. Appriciate your consistency - the KZbin algo does too ;)

[@vladimirbosinceanu5778]

I love your channel! Thank you so much!

[@Quell__]

just wanna say that you're my favourite youtuber right now, thank you for your content!

[@rpy9wyvbuerdopvpoe7blt77]

Hell yeah I've been looking for good in-depth AI news series

[@greyowlaudio]

We went from faking the moon landing to faking the moon.

[@danny3407]

I'm sorry, we can't discuss that at present. Thanks for your understanding.

[@ogredev]

@@danny3407 As an AI Language Model I am unable to reply to your response. However, it should be noted that girls rule and boys drool, and Trump is satan while Messiah Joe Biden is our savior.

[@malayariver]

All pictures of earth are fake as well

[@jcorey333]

Faking faking the moon landing was faked!

[@asdasd-yr7wi]

First KZbin to give decent sources :D good news, very nice! GOoooooooooooooooooooooooooooooooooOD job!

[@tristanwegner]

That such obvious improvements/low hanging fruits like restricting the physical search space to the correct units shows how early we still are with AI and how much discoveries are still waiting for those who want to look.

[@mohamedhatem325]

Another GREAT VIDEO, thanks

[@zakuro8532]

What a rumbling in AI research, and more so to come. Very inspirational.

[@3isr3g3n]

Dammisiech am i scared and excited at the same time. Also great video! This is certainly the best channel for AI news!

[@Aldraz]

There is so much AI news each and every day, it is crazy! We are truly in exponential phase. I am having problems to keep up with all of the new informations.

[@fyodorminakov6092]

Love the Curb bit.

[@alan2here]

It's (awful) response: With these ingredients, I recommend making Spicy Peanut Butter Couscous with Salami Chips and Chocolate Truffle Drizzle. Ingredients: - 1 cup couscous - 1 1/2 cups water - 1 tablespoon olive oil - 1/2 teaspoon sea salt - 1/4 cup chunky peanut butter - 1 tablespoon cooking oil - 1 tablespoon marmite - 1 tablespoon truffle marmite - 1 tablespoon sriracha - 1 tablespoon super hot chili sauce - 4 slices of salami - 1/4 cup 100% cocoa chocolate - 1 tablespoon cooking oil - 1 teaspoon white powdery stuff (could be flour or sugar) Instructions: 1. Preheat the oven to 375°F (190°C). Line a baking sheet with parchment paper. 2. In a medium saucepan, bring the water, olive oil, and sea salt to a boil. Stir in the couscous, cover, and remove from heat. Let it sit for 5 minutes, then fluff with a fork. 3. In a small bowl, mix together the peanut butter, marmite, truffle marmite, sriracha, and super hot.

[@JorgetePanete]

Its*

[@richarddanielzoom]

thank you, Yannic ,this epoch is

[@dreamphoenix]

Thank you.

[@fburton8]

FYI the moon’s aspect does move around a bit at any single geographical location, i.e. it’s not a fixed image. That doesn’t invalidate the point about what Samsung is alleged to be doing, of course.

[@angrydinosaur8plus9]

I saw on a video (with OpenAI employees I think it was) that they didn't want people to build up such high expectations for GPT4 because they weren't sure we'd feel it warranted it... something to that extent. So the (sort of lack of) announcement actually makes sense

[@byrnemeister2008]

And then BANG!

[@Hypersniper05]

I've been using gpt4 , BIG difference with coding. Much more accurate (not a 100%) , I can see the difference in chat memory as well. I use it over 3.5

[@nazar.shmatko]

Thank you for saving my time

[@morn1415]

Side note: The Moon is not tidally locked because of interaction with the water on earth, it is because of the the gravitational pull of the Earth on the Moon. The gravitational attraction of the Earth is stronger on the side of the Moon closer to us than on the side farther away. This creates a stretching force that tries to pull the Moon into an elongated shape. Over time the deformation has caused a bulge on the side of the Moon facing the Earth which creates an additional force that slows down the Moon's rotation, until the rotation rate and orbital period of the Moon become synchronized, so that the same side of the Moon always faces the Earth. ☝Which in the end makes it easier for Samsung to train an AI model... So are they involved with the creation of the Universe or the constants of Physics? Very suspicious.... 🤔

[@eelcohoogendoorn8044]

Wrt samsung; yeah that is exactly what is going on. In their defence, it is also the exact same thing as is going on with the supposed 'black hole image'. Nor is it just a thing with astronomy. I spent a large part of my time in grad school trying to explain to biologists that this is not ok and trying to explain to them the difference between image reconstruction based on conservative and legitimate inferences from data you have collected, versus inpainting stuff based on priors or models that are a complete black box and may be 100% inappropriate. I was not very successful. Astronomers and biologist will throw their money at the shiniest picture; and I am not going to hold my breath at the samsung-selfie crowd holding the market to different standards. I mean; its not even wrong in this case; its actually making the correct inference. It is the moon. It actually has that texture. Plus, it looks good on your insta. If you are going to be annoyed at this trend, phone pictures are not the right tree to bark up to.

[@videowatching9576]

Wow fascinating. The grounding things makes me think about various ways that editing tools can enable directly what you want. Curious how these things iterate on user experience, such as the example shown in the chat version for identifying the outlines, and perhaps making further adjustments to that. So some combination of text and UIs, with UI benefit being showing many possibilities. And perhaps text requests can even show how the UI would do it, for sake of helping to teach manual input and steps as well.

[@P1XeLIsNotALittleSquare]

On huggingface: most open source licenses have some sort of restrictions anyways

[@hartvenus]

No open source license is restricted access, that's stretching the definition a little too far lmao. It's a big jump from open-source by default to being able to put any arbitrary restriction on it, and i guarantee that in the next few years like 50% of the page will be paid only models

[@alan2here]

Bing AI now threatens you if you ask for lyrics interpretations! Unusually for this month GPT-3 gives a great reply and Bing nonsense doesn't. It's a very efficient way to make their customers incredibly angry, maybe if microshaft weren't such bags in the first place then they'd not need to be so paranoid about PR that they permanently lose customers and become irrelevant.

[@bart2019]

I can't help but wonder how you edit your monologues. At times it's really noticable that the video jumps every few words, but the audio from it flows very smoothly. For the latter: well done.

[@2DReanimation]

carykh the KZbinr created a Python script for detecting silence in videos and cutting out the silence defined by some threshold. That would be so helpful for a lot of KZbinrs, but I don't know if anyone knows of it.

[@jacobyoung6876]

I find it extremely distracting. It stutters too often and needs some audio & visual smoothing (or interpolation) in between edited clips.

[@AllExistence]

The fact that visual gpt has natural text instructions on how to answer to a human in the code is hilarious.

[@potts995]

Happy GANing!

[@alan2here]

Set the Bing app to "creative" and it seems to be more accurate than (optionally "chat") GPT-3 at reasoning about technical topics.

[@byrnemeister2008]

Because it’s GPT 4. Big step ahead on the reasoning front. In my experience so far.

[@Veptis]

Thank you for providing all these videos right now. I didn't have the motivation to read the 'techincal paper' marketing myself. But relied on your summary for talking about it in various places. The eval attempt to harness the open source community is sad. I am working on writing an evaluation suite for a dataset I am providing with all the Huggingface classes for it as part of a term paper. And I am still questioning myself - why doesn't OpenAI just write a wrapper for that. Especially after some employees on Twitter answered questions about it without answering the question. Does upscaling really 'invent' information? Since you started with a high res image and then blurred it with some kernal. And that means you included more information than if you started with a low resolution image to begin with. This also means, you could alter the moon to a bit and blur it. And get exactly the same results through your kernel.

[@Temari_Virus]

Think about it this way: when you create the low res image, there were other high res images you could've started with to get the same low res image. Since the model only knows about the low res image, it cannot know for sure which high res image was the original. So, even a "perfect" model must "invent" information to upscale the image. (like how if I tell you that x^2 = 1, you can't be sure if x = 1 or x = -1)

[@ahabkapitany]

I got to the point where I use Bing for 90% of my searches. Also, switched to Edge beta because of the Bing side panel integration.

[@latentessence]

"you mau recognize me from the video you are watching" damn nice catch

[@user-wr4yl7tx3w]

How do you keep up and cover so much materials?

[@andreassyren329]

The exponential growth in the field has been clear for a while. Two years ago I was making arguments for how we need to make big swings at my former employer or we'd fall behind within 6 months. Even so, literally making the argument that it is exponential and we will not be able to imagine the acceleration, I am completely in awe at what the models are doing now. Humans really are not equiped to imagine exponential growth.

[@alonamaloh]

A Gaussian blur doesn't actually remove information, since it is a full-rank linear mapping. The camera could easily have a processing step where it tries to find the best possible convolution to sharpen the image and that would explain everything in that Reddit post: The blurred image can easily be un-blurred, but if you make a sharp cut by showing it only half of the blurred image, it won't think the image is blurry because the bottom edge is sharp. Falsifiable prediction of my own: If you cut half the moon first and *then* apply Gaussian blur, it will un-blur it just fine. Another one: If you use a median filter instead of Gaussian blur, the phone won't be able to fix the image.

[@Regic]

While you are correct that guassian blur alone would not destroy information ('same' convolution, at least) in the reddit post the author claims he also downsized the image too which does. The post also has some other experiments which make it very clear that deconvolution is not the reason. They also found a description where samsung admits it uses AI to enhance images. It's probably a super resolution network.

[@alonamaloh]

@@Regic Ah, thank you so much for the clarification. Everything in Yannic's description of the situation seemed completely unremarkable to me.

[@sinarchbishop5102]

4:28 xD You are a god

[@igorpalecek3176]

Absolutely repsect Yannick for leaving the false prediction in! Love it. Interesting also to learn it is not the gradient of the gravity of the the whole earth that tidally locks the moon, but only the water!

[@Klokinator]

Yannic, what's happening with OpenAssistant? Any cool news on that front? I helped label a few hundred things so I'm hoping to see a big release soon!

[@JohnDlugosz]

You should have gone on to say that GTP-4 became available (text only) to ChatGTP Plus users with the update on the 14th (show OpenAI blog post), and the developer demo of multi-modal GTP-4 took place on whatever date, along with confirmation that Bing is using GTP-4 all along. Your own Full Analysis video was three days ago!

[@JorgetePanete]

gpt*

[@skierpage]

39:01 GILGEN rendered cars that are too long behind Walter White. The front and back are okay, but the obscured section means overall the car makes no sense. I tried the demo with "Walter White in Red Dead Redemption 2," and the AI generates a more realistic car if it isn't bisected by another generated elements in the foreground. This was another great ml news, but it was light on coverage of helpers and libraries. There's just too much going on!

[@rustybolts8953]

Rusty nails are great snacks. Plenty of iron. Cans of beans or mushy peas are even better if you open the can first.

[@acornstudios3824]

The Bing chatbot seems to be collecting data from users by asking them questions about themselves (scary), but unlike chat gpt, you can't really seem to ask it any questions, as it just closes the chat. Users be warned, I have no idea what they are doing with your data.

[@alexforget]

My prediction: a JARVIS assistant running this year on our computer (no cloud) with consumer GPU we already have.

[@cristianandrei5462]

Yeah, I've seen a recent development of Standford university where they used the self instruct data set and the llama model to reduce cost dramatically and still get a functioning model

[@noname-gp6hk]

That's the only way I would use one of these chatbots, is if it ran locally on my computer and answered to me instead of some company. I 100% want one as an AI Waifu but I'm not gonna get attached to one running in the cloud where it can just be shut off one day ala replika. I want one that lives on my device that I can take care of.

[@akirachisaka9997]

Personally I don’t really mind my phone doing weird super resolution shit and gives me pictures based on what the AI hallucinates. I mean, as long as the picture produced do resembles the picture I want to take pretty well.

[@gr8ape111]

Thats the beginning of the end of truth

[@noname-gp6hk]

@@gr8ape111 I'd say that ship passed a long time ago. It's already unrealistic to know the truth about nearly anything, except maybe subject matters which you have legit expertise in. We all kind of just relying on taking what other people tell us at face value. There's effectively no such thing as truth, it's effectively subjective and unknowable.

[@dinkusstinkus4396]

only for one week!! this week is 1000x as insane, have you seen the new study from microsoft??

[@sujitmourya]

Love this❤

[@2DReanimation]

carykh the KZbinr created a Python script for detecting silence in videos and cutting out the silence defined by some threshold. That would be so helpful for a lot of KZbinrs, but I don't know if anyone knows of it. I post this here in hopes of Yannic being able to optimize his editing, as he uses a lot of cuts.

[@patuko5304]

So anyone else wondering why Yanick didnt mentioned yet Alpaca Standford??Seems like its a huge deal that LLAMA 7B could perform as well as GPT3 by fine tune selfinstruct. Hope he digs into that!

[@danielbohringer1304]

yould you possibly review the languange model "rwkv"?

[@cherubin7th]

33:20 But if I download it, I didn't agree to any TOS of OpenAI.

[@MegaNightdude]

Yannick, you're absolutely the best.

[@christopheraaron2412]

Could something like physical symbolic optimization be combined with digital physics from Stephen Wolfram? Or are they already doing that?

[@neon_Nomad]

Hugging face was purchased by amazon

[@netscrooge]

I think it's time for us to launch a CERN-scale project to build an AI scientist with higher reasoning (GFlowNet, whatever) that can teach and advance itself by reviewing vast amounts of existing research, by utilizing a variety of existing math and science tools (Wolfram Mathematica, etc.), and by directly experimenting with supercomputer simulations. Your thoughts?

[@maheshsrocks]

🔥

[@DarkTrapStudio]

MKHD made a video on the samsung moon case and if I remember correctly Samsung answered

[@darkWolf354]

If you go to wiki... "ChatGPT-4 was released on March 14, 2023 and will be available via API and for premium ChatGPT users." ... like Microsoft? probably much more than a premium user!

[@and1play5]

u said no gpt 4 comes out this week? as its been out for days lmao

[@XOPOIIIO]

I think it's not a fake, Uruguay employees are saying the same in Spanish.

[@rkoll33]

Midjourney 5 Alpha is als out now

[@billykotsos4642]

bruh this week has been fucking insane

[@timmygilbert4102]

I feel the exponential

[@pengchengwu447]

The comment on the Anthropic AI safety blogpost is funny. And I don't hear any where mentioning Claude?

[@beecee793]

I got API access to GPT-4!

[@DannyCatacora]

Same! Yannic, let me know if you want to chat with it on my Discord server!

[@jasonbourne485]

The super resolution model doesn't feel like a full explanation. Look at the massive difference between the full moon and blurred half moon, I don't think a super resolution model would fully ignore that and there may be some deliberate moon recognition installed.

[@hartvenus]

Google is somehow managing to be the most impressive player in the AI space yet the worst one by far at releasing literally anything. No reason to trust them on how good the models are if they're not confident to even put up a waitlist. Maybe if Google wasn't a fat and lazy monopoly we'd have the AI utopia a few years earlier lol, but I still think that it's coming soon

[@alexforget]

Google was way too happy to sit on the tech and rake in the billions. Putting it out with their name on it is risky, people may start to understand how much they know about each one of us.

[@gr8ape111]

Google is done, its going to take a while, but watch their slooooooooow downfall over the next decade.

[@byrnemeister2008]

It’s the “innovators dilemma”. They have to much to lose for them to embrace AI. What’s the revenue model for AI? Paying $20 per month and competing with Open Ai is never going to replace their massive revenue from advertising. They are skewered.

[@khmnc]

12:30 no, the dark side of the moon is the side that is always facing away from the earth. it isn't called that because of it's light level but rather because no one knew what it looked like until recently. so the darkness refers to the level of knowledge rather then light.

[@JorgetePanete]

than*

[@hiiamlawliet480]

Gans are available since 2014 however some people say since 90's XDDDDDD Hello Sepp Hochreiter, Jürgen Schmidhuber

[@paulchatel2215]

I'm not in favor of putting pressure on meta about their Llama license. I'm just happy they released their model at all. Think about what OpenAI is doing. If Meta is getting bad press for actually releasing their model to the public, we know what they will do next time.

[@Verrisin]

32:30 - would a model for my company really count as "foundational or large scale" ?

[@Verrisin]

more and more I'm starting to "worry" that human cortex might be similar to an LLM ... surely not quite the same, but what if it turns out that it's ... maybe not equivalent - but at some level ... idk ... - Surely, we have a completely different performance characteristics and design, but I mean "functionally equivalent".

[@skierpage]

Large AI models ingest terabytes of information and somehow compress it into the weights of connections between orders of magnitude fewer artificial "neurons." They don't simply store the bytes and pixels of everything they ingest, yet they know a tremendous amount about their training data and use it when they generate new responses. The human brain _must_ be doing something similar.

[@thomasruhm1677]

How big is «an order of magnitude»? I hear this all the time.

[@user-wr4yl7tx3w]

I think the background drum effect may be distracting.

[@timokreuzer1820]

What if someone created a model, which has only the purpose of generating training data for other models, would that be considered a competing product? 🤔

[@billykotsos4642]

33:18 yh... but theres no way to tell what data a model has been trained on...

[@Verrisin]

21:20 - doesn't GPL put pretty heavy usage restrictions? There are often many things you have to ensure and comply with to use it, so ... - I agree that I don't like that feature much, but it's not so black and white to make it not open source ... I think?

[@shyrealist]

Came here to say this! Y'all gotta reread your GPL! Whereas, MIT licence is a very different kettle... But yes, I also agree with the general sentiment. The openness of recent years has been huge for research. If they begin to close then it means they've run out of ideas and they have to protect future revenues 🤷‍♂️

[@skierpage]

Oversimplifying, the GPL restricts your freedom to make a closed version of something released under GPL. Nerds have spent millions of man-hours arguing whether that makes it more or less free than open source licenses without that restriction. Many licenses have other restrictions on what you can do with something: you can't release it without the existing copyright and attribution, you can't make a commercial version, you can't use it for war, etc.

[@notnotandrew]

Is it possible that Samsung's upscaling model may have been trained on images of the moon and might have some information encoded in the model itself? Edit: I should have kept watching >_

[@hav0ur9O7]

Why do you have a green screen only not to use it?

[@quantumastrologer5599]

I love that smartphone producers decided to invalidate smart phone photography and give traditional photography room to grow again.

[@noname-gp6hk]

I don't know lad, seems like this just enhances cell phone cameras beyond what normal cameras can do. It's kind of like saying 'I sure am glad that these companies made forklifts so strong, it invalidates what they can lift and gives humans room to grow in the picking things up field again'. I'd just use an AI assisted cell phone to take pictures because the pictures will be better, just the same as I'd use a forklift to move things around a warehouse instead of people because the forklift performs better. I don't care how the work is done, all I care about is the output..

[@sevret313]

Can someone just force OpenAI to change name for ClosedAI now? When they don't even allow people to use their services to create training data it is becoming ridiculous.

[@robo1540]

why would they build an ai for that tho 😭applying a hd texture would 1 be more cost effective and 2 might just straight up look better

[@SimonJackson13]

The blur does nothing but filter into a gaussian. A convolution allows deblurring.

[@vslaykovsky]

You've been Schmidhubered in this video

[@Robubbabub]

Imagine a toaster becoming sentient and realizing what it's a part of 😭

[@segfault-]

Looks like my job is safe. These models still have a long way to go. Rusty nails and benzodiazepines are for sure edible.

[@Show__Dont__Tell]

Why aren't you covering the "Sparks in AGI" paper, its bizarre!!

[@astronemir]

Anyone have an AI summary of This weeks news in text form so I can search it? 😂